def generate_projectspace(self, geometry):
dtype = self.imagespace.dtype
proj_fspace = FunctionSpace(geometry.params, out_dtype=dtype)
if not self.imagespace.is_weighted:
weighting = None
elif (isinstance(self.imagespace.weighting, ConstWeighting)
and np.isclose(self.imagespace.weighting.const,
self.imagespace.cell_volume)):
extent = float(geometry.partition.extent.prod())
size = float(geometry.partition.size)
weighting = extent / size
else:
raise NotImplementedError('unknown weighting of domain')
proj_tspace = self.imagespace.tspace_type(geometry.partition.shape,
weighting=weighting,
dtype=dtype)
if geometry.motion_partition.ndim == 0:
angle_labels = []
if geometry.motion_partition.ndim == 1:
angle_labels = ['$\\varphi$']
elif geometry.motion_partition.ndim == 2:
# TODO: check order
angle_labels = ['$\\vartheta$', '$\\varphi$']
elif geometry.motion_partition.ndim == 3:
# TODO: check order
angle_labels = ['$\\vartheta$', '$\\varphi$', '$\\psi$']
else:
angle_labels = None
if geometry.det_partition.ndim == 1:
det_labels = ['$s$']
elif geometry.det_partition.ndim == 2:
det_labels = ['$u$', '$v$']
else:
det_labels = None
if angle_labels is None or det_labels is None:
# Fallback for unknown configuration
axis_labels = None
else:
axis_labels = angle_labels + det_labels
proj_interp = 'nearest'
# proj_interp = kwargs.get('interp', 'nearest')
return DiscreteLp(proj_fspace,
geometry.partition,
proj_tspace,
interp=proj_interp,
axis_labels=axis_labels)
def _resize_discr(discr, newshp, offset, discr_kwargs):
"""Return a space based on ``discr`` and ``newshp``.
Use the domain of ``discr`` and its partition to create a new
uniformly discretized space with ``newshp`` as shape. In axes where
``offset`` is given, it determines the number of added/removed cells to
the left. Where ``offset`` is ``None``, the points are distributed
evenly to left and right. The ``discr_kwargs`` parameter is passed
to `uniform_discr` for further specification of discretization
parameters.
"""
nodes_on_bdry = discr_kwargs.get('nodes_on_bdry', False)
if np.shape(nodes_on_bdry) == ():
nodes_on_bdry = ([(bool(nodes_on_bdry), bool(nodes_on_bdry))] *
discr.ndim)
elif discr.ndim == 1 and len(nodes_on_bdry) == 2:
nodes_on_bdry = [nodes_on_bdry]
elif len(nodes_on_bdry) != discr.ndim:
raise ValueError('`nodes_on_bdry` has length {}, expected {}'
''.format(len(nodes_on_bdry), discr.ndim))
dtype = discr_kwargs.pop('dtype', discr.dtype)
impl = discr_kwargs.pop('impl', discr.impl)
exponent = discr_kwargs.pop('exponent', discr.exponent)
interp = discr_kwargs.pop('interp', discr.interp)
weighting = discr_kwargs.pop('weighting', discr.weighting)
affected = np.not_equal(newshp, discr.shape)
ndim = discr.ndim
for i in range(ndim):
if affected[i] and not discr.is_uniform_byaxis[i]:
raise ValueError('cannot resize in non-uniformly discretized '
'axis {}'.format(i))
grid_min, grid_max = discr.grid.min(), discr.grid.max()
cell_size = discr.cell_sides
new_minpt, new_maxpt = [], []
for axis, (n_orig, n_new, off, on_bdry) in enumerate(
zip(discr.shape, newshp, offset, nodes_on_bdry)):
if not affected[axis]:
new_minpt.append(discr.min_pt[axis])
new_maxpt.append(discr.max_pt[axis])
continue
n_diff = n_new - n_orig
if off is None:
num_r = n_diff // 2
num_l = n_diff - num_r
else:
num_r = n_diff - off
num_l = off
try:
on_bdry_l, on_bdry_r = on_bdry
except TypeError:
on_bdry_l = on_bdry
on_bdry_r = on_bdry
if on_bdry_l:
new_minpt.append(grid_min[axis] - num_l * cell_size[axis])
else:
new_minpt.append(grid_min[axis] - (num_l + 0.5) * cell_size[axis])
if on_bdry_r:
new_maxpt.append(grid_max[axis] + num_r * cell_size[axis])
else:
new_maxpt.append(grid_max[axis] + (num_r + 0.5) * cell_size[axis])
fspace = FunctionSpace(IntervalProd(new_minpt, new_maxpt), out_dtype=dtype)
tspace = tensor_space(newshp,
dtype=dtype,
impl=impl,
exponent=exponent,
weighting=weighting)
# Stack together the (unchanged) nonuniform axes and the (new) uniform
# axes in the right order
part = uniform_partition([], [], ())
for i in range(ndim):
if discr.is_uniform_byaxis[i]:
part = part.append(
uniform_partition(new_minpt[i],
new_maxpt[i],
newshp[i],
nodes_on_bdry=nodes_on_bdry[i]))
else:
part = part.append(discr.partition.byaxis[i])
return DiscreteLp(fspace, part, tspace, interp=interp)
def __init__(self, reco_space, geometry, variant, **kwargs):
"""Initialize a new instance.
Parameters
----------
reco_space : `DiscreteLp`
Discretized reconstruction space, the domain of the forward
operator or the range of the adjoint (back-projection).
geometry : `Geometry`
Geometry of the transform that contains information about
the data structure.
variant : {'forward', 'backward'}
Variant of the transform, i.e., whether the ray transform
or its back-projection should be created.
Other Parameters
----------------
impl : {`None`, 'astra_cuda', 'astra_cpu', 'skimage'}, optional
Implementation back-end for the transform. Supported back-ends:
- ``'astra_cuda'``: ASTRA toolbox, using CUDA, 2D or 3D
- ``'astra_cpu'``: ASTRA toolbox using CPU, only 2D
- ``'skimage'``: scikit-image, only 2D parallel with square
reconstruction space.
For the default ``None``, the fastest available back-end is
used.
interp : {'nearest', 'linear'}, optional
Interpolation type for the discretization of the projection
space. This has no effect if ``proj_space`` is given explicitly.
Default: ``'nearest'``
proj_space : `DiscreteLp`, optional
Discretized projection (sinogram) space, the range of the forward
operator or the domain of the adjoint (back-projection).
Default: Inferred from parameters.
use_cache : bool, optional
If ``True``, data is cached. This gives a significant speed-up
at the expense of a notable memory overhead, both on the GPU
and on the CPU, since a full volume and a projection dataset
are stored. That may be prohibitive in 3D.
Default: True
Notes
-----
The ASTRA backend is faster if data are given with
``dtype='float32'`` and storage order 'C'. Otherwise copies will be
needed.
"""
variant, variant_in = str(variant).lower(), variant
if variant not in ('forward', 'backward'):
raise ValueError('`variant` {!r} not understood'
''.format(variant_in))
if variant == 'forward':
reco_name = 'domain'
proj_name = 'range'
else:
reco_name = 'range'
proj_name = 'domain'
if not isinstance(reco_space, DiscreteLp):
raise TypeError('`{}` must be a `DiscreteLp` instance, got '
'{!r}'.format(reco_name, reco_space))
if not isinstance(geometry, Geometry):
raise TypeError('`geometry` must be a `Geometry` instance, got '
'{!r}'.format(geometry))
# Handle backend choice
if not _AVAILABLE_IMPLS:
raise RuntimeError('no ray transform back-end available; '
'this requires 3rd party packages, please '
'check the install docs')
impl = kwargs.pop('impl', None)
if impl is None:
# Select fastest available
if ASTRA_CUDA_AVAILABLE:
impl = 'astra_cuda'
elif ASTRA_AVAILABLE:
impl = 'astra_cpu'
elif SKIMAGE_AVAILABLE:
impl = 'skimage'
else:
raise RuntimeError('bad impl')
else:
impl, impl_in = str(impl).lower(), impl
if impl not in _SUPPORTED_IMPL:
raise ValueError('`impl` {!r} not understood'.format(impl_in))
if impl not in _AVAILABLE_IMPLS:
raise ValueError('{!r} back-end not available'.format(impl))
# Cache for input/output arrays of transforms
self.use_cache = kwargs.pop('use_cache', True)
# Sanity checks
if impl.startswith('astra'):
if geometry.ndim > 2 and impl.endswith('cpu'):
raise ValueError('`impl` {!r} only works for 2d'
''.format(impl_in))
# Print a warning if the detector midpoint normal vector at any
# angle is perpendicular to the geometry axis in parallel 3d
# single-axis geometry -- this is broken in some ASTRA versions
if (isinstance(geometry, Parallel3dAxisGeometry)
and not astra_supports('par3d_det_mid_pt_perp_to_axis')):
axis = geometry.axis
mid_pt = geometry.det_params.mid_pt
for i, angle in enumerate(geometry.angles):
if abs(np.dot(axis, geometry.det_to_src(angle,
mid_pt))) < 1e-4:
warnings.warn(
'angle {}: detector midpoint normal {} is '
'perpendicular to the geometry axis {} in '
'`Parallel3dAxisGeometry`; this is broken in '
'ASTRA v{}, please upgrade to v1.8 or later'
''.format(i, geometry.det_to_src(angle, mid_pt),
axis, ASTRA_VERSION), RuntimeWarning)
break
elif impl == 'skimage':
if not isinstance(geometry, Parallel2dGeometry):
raise TypeError("{!r} backend only supports 2d parallel "
'geometries'.format(impl))
mid_pt = reco_space.domain.mid_pt
if not np.allclose(mid_pt, [0, 0]):
raise ValueError('`{}` must be centered at (0, 0), got '
'midpoint {}'.format(reco_name, mid_pt))
shape = reco_space.shape
if shape[0] != shape[1]:
raise ValueError('`{}.shape` must have equal entries, '
'got {}'.format(reco_name, shape))
extent = reco_space.domain.extent
if extent[0] != extent[1]:
raise ValueError('`{}.extent` must have equal entries, '
'got {}'.format(reco_name, extent))
if reco_space.ndim != geometry.ndim:
raise ValueError('`{}.ndim` not equal to `geometry.ndim`: '
'{} != {}'.format(reco_name, reco_space.ndim,
geometry.ndim))
self.__geometry = geometry
self.__impl = impl
# Generate or check projection space
proj_space = kwargs.pop('proj_space', None)
if proj_space is None:
dtype = reco_space.dtype
proj_uspace = FunctionSpace(geometry.params, out_dtype=dtype)
if isinstance(reco_space.weighting, NoWeighting):
weighting = 1.0
elif (isinstance(reco_space.weighting, ConstWeighting)
and np.isclose(reco_space.weighting.const,
reco_space.cell_volume)):
# Approximate cell volume
# TODO: find a way to treat angles and detector differently
# regarding weighting. While the detector should be uniformly
# discretized, the angles do not have to and often are not.
# The needed partition property is available since
# commit a551190d, but weighting is not adapted yet.
# See also issue #286
extent = float(geometry.partition.extent.prod())
size = float(geometry.partition.size)
weighting = extent / size
else:
raise NotImplementedError('unknown weighting of domain')
proj_dspace = reco_space.dspace_type(geometry.partition.size,
weighting=weighting,
dtype=dtype)
if geometry.ndim == 2:
axis_labels = ['$\\theta$', '$s$']
elif geometry.ndim == 3:
axis_labels = ['$\\theta$', '$u$', '$v$']
else:
# TODO Add this when we add nd ray transform
axis_labels = None
proj_interp = kwargs.get('interp', 'nearest')
proj_space = DiscreteLp(proj_uspace,
geometry.partition,
proj_dspace,
interp=proj_interp,
order=reco_space.order,
axis_labels=axis_labels)
else:
# proj_space was given, checking some stuff
if not isinstance(proj_space, DiscreteLp):
raise TypeError('`{}` must be a `DiscreteLp` instance, '
'got {!r}'.format(proj_name, proj_space))
if proj_space.shape != geometry.partition.shape:
raise ValueError('`{}.shape` not equal to `geometry.shape`: '
'{} != {}'.format(proj_name, proj_space.shape,
geometry.partition.shape))
if proj_space.dtype != reco_space.dtype:
raise ValueError('`{}.dtype` not equal to `{}.dtype`: '
'{} != {}'.format(proj_name, reco_name,
proj_space.dtype,
reco_space.dtype))
# Reserve name for cached properties (used for efficiency reasons)
self._adjoint = None
self._astra_wrapper = None
# Extra kwargs that can be reused for adjoint etc. These must
# be retrieved with `get` instead of `pop` above.
self._extra_kwargs = kwargs
# Finally, initialize the Operator structure
if variant == 'forward':
super().__init__(domain=reco_space, range=proj_space, linear=True)
elif variant == 'backward':
super().__init__(domain=proj_space, range=reco_space, linear=True)
def __init__(self, discr_domain, geometry, **kwargs):
"""Initialize a new instance.
Parameters
----------
discr_domain : `DiscreteLp`
Discretized space, the domain of the forward projector
geometry : `Geometry`
Geometry of the transform, containing information about
the operator range
Other Parameters
----------------
impl : {`None`, 'astra_cuda', 'astra_cpu', 'scikit'}, optional
Implementation back-end for the transform. Supported back-ends:
* ``'astra_cuda'``: ASTRA toolbox, using CUDA, 2D or 3D
* ``'astra_cpu'``: ASTRA toolbox using CPU, only 2D
* ``'scikit'``: scikit-image, only 2D parallel with square domain
If ``None`` is given, the fastest available back-end is used.
interp : {'nearest', 'linear'}, optional
Interpolation type for the discretization of the operator
range.
Default: 'nearest'
discr_range : `DiscreteLp`, optional
Discretized space, the range of the forward projector.
Default: Infered from parameters.
use_cache : bool, optional
If ``True``, data is cached. Note that this causes notable memory
overhead, both on the GPU and on the CPU since a full volume and
projection is stored. In the 3D case, some users may want to
disable this.
Default: True
Notes
-----
The ASTRA backend is faster if data is given with ``dtype`` 'float32'
and storage order 'C'. Otherwise copies will be needed.
"""
if not isinstance(discr_domain, DiscreteLp):
raise TypeError('`discr_domain` {!r} is not a `DiscreteLp`'
' instance'.format(discr_domain))
if not isinstance(geometry, Geometry):
raise TypeError('`geometry` {!r} is not a `Geometry` instance'
''.format(geometry))
impl = kwargs.pop('impl', None)
if impl is None:
# Select fastest available
if ASTRA_CUDA_AVAILABLE:
impl = 'astra_cuda'
elif ASTRA_AVAILABLE:
impl = 'astra_cpu'
elif SCIKIT_IMAGE_AVAILABLE:
impl = 'scikit'
else:
raise ValueError('no valid `impl` installed')
impl, impl_in = str(impl).lower(), impl
if impl not in _SUPPORTED_IMPL:
raise ValueError('`impl` {!r} not supported' ''.format(impl_in))
self.use_cache = kwargs.pop('use_cache', True)
# TODO: sanity checks between impl and discretization impl
if impl.startswith('astra'):
# TODO: these should be moved somewhere else
if not ASTRA_AVAILABLE:
raise ValueError("'astra' back-end not available")
if impl == 'astra_cuda' and not ASTRA_CUDA_AVAILABLE:
raise ValueError("'astra_cuda' back-end not available")
if not np.allclose(discr_domain.partition.cell_sides[1:],
discr_domain.partition.cell_sides[:-1]):
raise ValueError('ASTRA does not support different voxel '
'sizes per axis, got {}'
''.format(discr_domain.partition.cell_sides))
if geometry.ndim > 2 and impl.endswith('cpu'):
raise ValueError('`impl` {}, only works for 2d geometries'
' got {}-d'.format(impl_in, geometry))
elif impl == 'scikit':
if not isinstance(geometry, Parallel2dGeometry):
raise TypeError("'scikit' backend only supports 2d parallel "
'geometries')
mid_pt = discr_domain.domain.mid_pt
if not all(mid_pt == [0, 0]):
raise ValueError('`discr_domain.domain` needs to be '
'centered on [0, 0], got {}'.format(mid_pt))
shape = discr_domain.shape
if shape[0] != shape[1]:
raise ValueError('`discr_domain.shape` needs to be square '
'got {}'.format(shape))
extent = discr_domain.domain.extent
if extent[0] != extent[1]:
raise ValueError('`discr_domain.extent` needs to be square '
'got {}'.format(extent))
# TODO: sanity checks between domain and geometry (ndim, ...)
self.__geometry = geometry
self.__impl = impl
self.kwargs = kwargs
discr_range = kwargs.pop('discr_range', None)
if discr_range is None:
dtype = discr_domain.dspace.dtype
# Create a discretized space (operator range) with the same
# data-space type as the domain.
# TODO: use a ProductSpace structure or find a way to treat
# different dimensions differently in DiscreteLp
# (i.e. in partitions).
range_uspace = FunctionSpace(geometry.params, out_dtype=dtype)
# Approximate cell volume
# TODO: angles and detector must be handled separately. While the
# detector should be uniformly discretized, the angles do not have
# to and often are not.
extent = float(geometry.partition.extent.prod())
size = float(geometry.partition.size)
weight = extent / size
range_dspace = discr_domain.dspace_type(geometry.partition.size,
weighting=weight,
dtype=dtype)
if geometry.ndim == 2:
axis_labels = ['$\\theta$', '$s$']
elif geometry.ndim == 3:
axis_labels = ['$\\theta$', '$u$', '$v$']
else:
# TODO Add this when we add nd ray transform.
axis_labels = None
range_interp = kwargs.get('interp', 'nearest')
discr_range = DiscreteLp(range_uspace,
geometry.partition,
range_dspace,
interp=range_interp,
order=discr_domain.order,
axis_labels=axis_labels)
self.backproj = None
super().__init__(discr_domain, discr_range, linear=True)
def __init__(self, discr_range, geometry, **kwargs):
"""Initialize a new instance.
Parameters
----------
discr_range : `DiscreteLp`
Reconstruction space, the range of the back-projector
geometry : `Geometry`
The geometry of the transform, contains information about
the operator domain
Other Parameters
----------------
impl : {'astra_cpu', 'astra_cuda', 'scikit'}, optional
Implementation back-end for the transform. Supported back-ends:
* ``'astra_cuda'``: ASTRA toolbox, using CUDA, 2D or 3D
* ``'astra_cpu'``: ASTRA toolbox using CPU, only 2D
* ``'scikit'``: scikit-image, only 2D parallel with square domain
If ``None`` is given, the fastest available back-end is used.
interp : {'nearest', 'linear'}, optional
Interpolation type for the discretization of the operator range.
Default: 'nearest'
discr_domain : `DiscreteLp`, optional
Discretized space, the range of the forward projector.
Default: Infered from parameters.
use_cache : bool, optional
If ``True``, data is cached. Note that this causes notable memory
overhead, both on the GPU and on the CPU since a full volume and
projection is stored. In the 3D case, some users may want to
disable this.
"""
if not isinstance(discr_range, DiscreteLp):
raise TypeError('`discr_range` {!r} is not a `DiscreteLp`'
' instance'.format(discr_range))
if not isinstance(geometry, Geometry):
raise TypeError('`geometry` {!r} is not a `Geometry` instance'
''.format(geometry))
impl = kwargs.pop('impl', None)
if impl is None:
# Select fastest available
if ASTRA_CUDA_AVAILABLE:
impl = 'astra_cuda'
elif ASTRA_AVAILABLE:
impl = 'astra_cpu'
elif SCIKIT_IMAGE_AVAILABLE:
impl = 'scikit'
else:
raise ValueError('no valid `impl` installed')
impl, impl_in = str(impl).lower(), impl
if impl not in _SUPPORTED_IMPL:
raise ValueError("`impl` '{}' not supported" ''.format(impl_in))
if impl.startswith('astra'):
if not ASTRA_AVAILABLE:
raise ValueError("'astra' backend not available")
if impl == 'astra_cuda' and not ASTRA_CUDA_AVAILABLE:
raise ValueError("'astra_cuda' backend not available")
if not np.allclose(discr_range.partition.cell_sides[1:],
discr_range.partition.cell_sides[:-1],
atol=0,
rtol=1e-5):
raise ValueError('ASTRA does not support different voxel '
'sizes per axis, got {}'
''.format(discr_range.partition.cell_sides))
self.use_cache = kwargs.pop('use_cache', True)
self.__geometry = geometry
self.__impl = impl
self.kwargs = kwargs
discr_domain = kwargs.pop('discr_domain', None)
if discr_domain is None:
dtype = discr_range.dspace.dtype
# Create a discretized space (operator domain) with the same
# data-space type as the range.
domain_uspace = FunctionSpace(geometry.params, out_dtype=dtype)
# Approximate cell volume
extent = float(geometry.partition.extent.prod())
size = float(geometry.partition.size)
weight = extent / size
domain_dspace = discr_range.dspace_type(geometry.partition.size,
weighting=weight,
dtype=dtype)
if geometry.ndim == 2:
axis_labels = ['$\\theta$', '$s$']
elif geometry.ndim == 3:
axis_labels = ['$\\theta$', '$u$', '$v$']
else:
# TODO Add this when we add nd ray transform.
axis_labels = None
domain_interp = kwargs.get('interp', 'nearest')
discr_domain = DiscreteLp(domain_uspace,
geometry.partition,
domain_dspace,
interp=domain_interp,
order=discr_range.order,
axis_labels=axis_labels)
self.ray_trafo = None
super().__init__(discr_domain, discr_range, linear=True)